Will Meta’s AI Layoffs Backfire?

Hey, what's going on everybody? Welcome to The Daily AI Show. Today is May 19, 2026 and with me today we've got Andy, Beth, Brian. I was just realizing as I was getting going to this, we go on mute when you guys see the flash card, the open screen there. And I was having to clear my throat because you realize sometimes you haven't been talking to people in quite some time in the morning. And then I tried to talk to you guys and I was like, oh, that won't do it all for a while. So I was like, quickly clearing out the throat there. Hey Jeff, hey Cisco in the comments, I appreciate y'all being here. Yeah, I've kind of been heads down this morning and just working on some stuff. I know it's earlier where you are, Andy, me and Beth get the benefit of having a few hours. Although, Beth, you're a bit of a night hour, so I don't know how many hours you get prior to this show. You might be a little more like Andy as far as coming in, but I'm typically up at about 630. So I do have a few hours before the 10 o'clock hour rolls around. And I'm working on stuff. So let's see. Well, yesterday we kicked off with Beth. So why don't we switch it up? And Andy, let's kick it off with you with some of the news. I have some things to share on a couple of different topics as well as a show and tell. But Andy, let's get it going with you first. All right. Yeah, just the calendar reminder that 10 a.m. Pacific time, so three hours from now is the keynote of the Google I.O. conference. And so we didn't speak about it before, but maybe we'll do a live, live listen to it. I don't know if anybody's available to do that. But if we do, we'll let you know. And in the community, at least. And we'll try to watch it simultaneously together so we can kind of cast doubt on inflated announcements or go ooh and ah when things are really, really surprising. And that is a reason to hit the bell because if you hit the bell, you get notified every time we go live. But very good. And I was like to have really quick that because I was looking at the schedule for I.O. There is the keynote, obviously, as you just said, Andy. But for anybody that's interested, you can go and sign up and see the lives at I.O.Google forward slash 2026. That's the website. But anyway, at 3.30 Pacific, so 6.30 Eastern, that's specifically when they have the What's New in Google AI. They have several tracks, so at the same exact time, they also have What's New in Android, What's New in Chrome, which I would prefer, I would actually want to see as well. And they also have Agent First Workflows from production. So they have a lot going on. You're not going to be able to catch all of it super live, but for folks wanting to do that and wondering when those things happen beyond the keynote, as well as many other things they have going on. But I know for our audience, What's New in Google AI is probably going to be one of the first ones we care about. Of course, if you don't catch any of that, we will, and we will talk about it on tomorrow's show. So don't worry about it, even if you don't have time for it. Just tune in tomorrow for our Wednesday show and Thursday when we'll talk about both days of Google I.O. They just happen, obviously, after we're live. So we'll catch them on the next show. So all good things coming there. Andy, back to you. Yeah, so I think the news item that I want to talk about at the top of the show is that tomorrow, Meta is just throwing a hand grenade into their entire organization. And they have an overall reorganization that's focused on AI pods underway. They're going to riff, reduction in force, 8,000 employees at Meta. And it's going to happen in three waves. It's already happened. Well, no, I'm sorry. Today's Tuesday. Sorry. Wednesday morning, tomorrow at like 4 a.m., they're sending out the first round of emails to notify which Meta employees are being riffed. And then an hour later, another email goes out and sort of the second wave happens and then the third wave. So imagine how disoriented and disruptive that is to an organization that's working today and they're waiting tomorrow for the bomb to drop. So the HR chief at Meta, her name is Janelle Gale, posted a memo yesterday to an employee resource group that outlined which employees should expect. No, I'm sorry. They didn't say which. They said what employees should expect. And that's the revelation that they're going to do this in these three waves of emails notifying you if your job is gone. And then she also said that management positions would be cut across the company to create flatter organizational structures. And she said now we're at the stage where many organizations can operate with a flatter structure, smaller teams in AI pods that can move faster and with more ownership over those products. And she said they'll also move. So this is part of the reorganization. They'll move. I don't mean physical relocation, but they'll move in terms of who reports to whom. They're going to move 7,000 people in addition to the 8,000 they're cutting to work on new initiatives around AI that are crucial to Meta's forward-looking strategy. Yeah. So all of that in the context of a Gartner study that just was released that said that AI layoffs, and you can say that Meta's 8,000 reduction in force is going to be an AI-driven layoff, are backfiring. That 80% of companies that cut staff due to AI don't get an ROI that follows when they do that. And so they surveyed 350 global enterprises with revenue above 1 billion. So these are major enterprises and found that 80% had cut staff after AI deployments. And the uncomfortable finding was that companies that reduced that headcount were just as likely to see negative outcomes as they were to see positive ones. So, you know, is it really all that smart? Companies that actually have succeeded with ROI in the context of AI transformation are the ones that empowered employees using AI without doing massive rifts. So that's the opposite, right? Let's give our employees the tools that make them more effective, not just go after massive reductions in cost that are driven by concerns about the massive investments required to buy the compute to support AI. The other way around is that you have to basically have to swallow both at the same time. You got to keep those employees, give them the tools, and spend the additional money in order to transform using AI. Well, I want to talk about the Gartner Report because that was one of the big things I didn't get to yesterday. And I think there's more to it. But also just because I didn't know, I wanted to look up really quickly. Google says, Google AI says that as of March of 2026, Meta Platform's total workforce stood at 77, so let's just say 78,000. So you were just talking about the rifts there. That's a sizable percentage, 8,000 from that, as well as reassigning the 7,000. And so what I wonder at Meta, I've known people personally who've worked at Meta. I won't name any names or whatever. This is in years past, not necessarily in the era of AI, but in years past. And I get the impression, this is just me listening to them, that it's always been a little crazy at Meta. It's always been, if you work at Meta, there's a lot of reasons, the good reasons why you might professionally do that. It could be a great resume builder, potentially, for what you're trying to do. But oftentimes, you don't hear people staying there for super long periods of time because almost like a Tesla or otherwise, it could take a toll on your family as well as other things. At least these are some of the stories I would hear. That doesn't mean you shouldn't do it or you shouldn't work there, whatever. But maybe some people have amazing tenure at Meta, and I haven't heard those stories. Regardless, I do feel like we continuously keep hearing these same sort of reorg stories. And honestly, there was one with Greg Brockman at OpenAI yesterday, so I'm certainly not singling out Meta. But they seem to be in the news a lot about this sort of what feels like these monumental shifts, these monumental shifts. And this brings me back over to the other news report you were just talking about, Andy. Sorry, I have it here. Just give me a second. The CNBC, this is out of CNBC. Oh, I'm sorry. So this is from CNBC. Compiled a list, AI-related layoffs, a boost for stocks? Not necessarily, right? That came out on Sunday. And what I wanted to talk about on that is some of the stats that you talked about, Andy. They call out a couple of companies such as Salesforce, Nike, as well as Fiverr, as some of the companies they called out. And they have graphs that sort of show where the stock has moved since these major announcements for rifts that were AI-related. Now, a couple of things. CNBC points out that, is this just a cover for what would normally be another reason for job layoffs or whatever? And it's being attributed to AI. But maybe behind the scenes, it isn't. And I wonder that with Meta as well. You know, you see these reduction in forces or whatever. Is it just that Meta has been around so long that they've gotten bloat in a lot of areas? I mean, I certainly have seen this in corporations, not necessarily the ones. The areas I see usually are in government because of duplication of effort. I saw that all the time when I was in Iraq. Multiple people who essentially had the same job role. But that's government. You see that a lot. Well, with companies, is this... I guess what I'm getting at back to you guys. I want to try to be clear here. When I read something like this, I don't deny that there is a correlation between announcement made at Nike, stock price today. Can't deny that, right? But my question is, did Nike already... If we took iI Nike, did Nike already have other problems that they were dealing with before that announcement was made? That may also be more related to that stock decline. And could you go company by company, Salesforce and otherwise, and say, hey, correlation doesn't equal... Doesn't necessarily equal causation. And so I do wonder about that. That we see a story like this. And basically, they're pointing out and saying, hey, I mean, here are the facts. The stats don't lie on this. However, I wonder if it's just a bigger story than AI. Because of other reasons, many, many complicated reasons why companies might need to do rips or otherwise. And I feel like Meta is in that case too. What worries me with Meta, not that I don't have stock with them or anything, but worries me with Meta is that they seem to be doing this a lot lately in the last two, three years. There almost always seems to be this turnover. I don't know if the age of AI is causing that and causing companies to have to turn quickly and create new models. I agree with the AI pod idea. That's actually right. I mean, conceptually, that's right to me. That's how a company should be. Should be leaner, a little bit flatter. AI pods, move faster, be more nimble. And AI does allow you to do that. So conceptually, I listen to that, Andy, and I go, yeah, that sounds right. That sounds directionally right from little O'Brien. I don't run. I'm not Zuckerberg. I don't run the company. But that sounds correct. Glenn agrees. But I guess I'm not dismissing the AI or that the AI announcements at Fiverr didn't have a direct impact on stock. I'm not saying that at all. I'm just saying this is an interesting story, but is it really the whole story? To me, I bet you there's a lot more nuance. And you would really have to go company by company. Look at Salesforce and go, is AI, the announcement of job cuts due to AI, truly the reason for this stock slip? Or is it just something that happened to happen? I guess. I'm not an expert here. I just, I'm curious. Andy, you're muted. Let me just say that clearly there are potentially other beneficial reasons why you would want to do a reorganization and why you would cut staff at that scale. It's quite easy to have little pockets of low productivity per worker and misalignment with what your real strategic objectives are. So there's definitely other reasons than AI. But in the case of Meta and others, the huge numbers, the hundreds of billions of dollars that they're having to spend on AI infrastructure has to be offset by cuts in other expenses. And it's happening. So whether or not AI is the rationale for it, at the CFO level, the CFO is saying, look, we can't do both. We can't have all these people and also invest in all of this hardware. That capital expenditure has got to be offset by some other operational cost improvement. And so that's, I think, what's driving it at the CFO level. And one of the mentions in the Gartner report was that, I guess, they also surveyed CFOs about what their forward-looking plans are for AI-driven cuts. And the line was, you know, CFOs privately admit that it's going to be 10x what is publicly stated. We're actually cutting deeper than what we're announcing as AI-related because of AI investment. So I think a lot of this is not about operational efficiency, but there might be a sort of a knock-on benefit of improved operational efficiency as companies make a really hard, hard decision, but also are taking a hard look at where do we really get rubber meeting the road with employee investment. Right, because there may be some nuance about the ROI of the actual cuts, but the ROI of announcing the cuts has been shown to be negative, right? Like when you announce these cuts and everybody knows that that's what you're doing, people get very, very upset. So we're not going to do that, but we are going to do this little bit with an announcement. I also wonder, is it tricky, Andy, to have a company where your profit is not based on the AI part, right? Meta's profit is very much based on the social media tools that they offer and that people use. And they've been using machine learning for a long time. They have AI under the hood. It's not been like generative AI necessarily, but now we layer generative on top of the machine learning that was there. Yeah. Yeah, and I want to call out too, just from the CNBC article, as a counterpoint to your sales law, Fiverr, Nike, Alphabet, head parent company of Google. So the investor from here said, he cited Google, which is owned public by Alphabet, as I said, as an example of a firm that is boosting its business with AI. Its generative AI tool, Gemini, has contributed to cloud revenue, strengthened search, and boosted user engagement across the Google ecosystem. Now, that ties in nicely to Google I.O. And what we'll be talking about, we'll be talking a lot about Google the next few days. But their stock doesn't lie, right? When they're using, they've invested, and I imagine Meta very much wants to be doing this as well with Meta AI into WhatsApp and all the other areas they work. They're investing and then trying to make sure that their AI is improving and improving the user experience overall as well as the tools. But as of right now, Google stock is not only up this year, beginning of the year, it was at about 317. It's now sitting at 390. But then if you go out to a one year, we're talking about 168 to 390 in one year and five year similar. They had a drop in 2023 where they dropped a bit. They were in 2021-23. They dropped down to as low as 86 at one point in probably, it looks like early 23. And then it's a bit of a roller coaster, and they've had some dips in there in early 25. But there's no doubt about it. It's up until the right right now. Whether it's going to stay that way or not, I don't know. Well, I think that's a growing recognition that Google is going to be the survivor in the long run. And it's a safer bet in AI than others. Yeah. I mean, that's probably what people see in NVIDIA as well. They're just not a one-stop shop, one-trick pony. Either of them. We'll take both. I'll take both. Yeah, those are two opposing ideas. But yeah. It's a one-pony shop here. We're talking about the one-pony shop. And I have a shop. NVIDIA, yeah. NVIDIA has many ponies in the stable. And I said we keep going with this. Pre-cub was just around, right? So we keep talking about horses and pony. This is kind of a related point. I read this morning, and I don't have a reference for it, but I read that at a couple of college commencement speeches, when the speakers, you know, the commencement speakers, like famous people or, you know, industry representatives speak to the graduating class. When they praised AI and did that, they got booed. And that the new generations, you can imagine, aren't real happy about what AI is doing to their future prospects. Well, they just invested a ton of money in college. And AI is taking not only the job opportunity at the low end of the rungs of the ladder that they were planning to step on, but they're actually firing huge numbers of people who are going to now be competing with those people. And it's just collapsing their promise of, you know, a rich and bright future. But, you know, of course, many of the commencement speakers are going to be saying, your future is so bright, you got to wear shades. Well, look, two things here. One, I got to tell you about the best commencement speech I've seen in quite some time. I just happened this season. So I'll say that in a second. But two, imagine that. Imagine you're a parent in the audience or you're actually the student. And four years ago, you made a good decision with all your math background to be a coder just four years ago. And now here you are. And even two years ago, you still felt confident enough to take that on as a major. And then you went through all the stuff. Now, is that person not going to be able to get a job? I mean, entry-level coder could be tough. Could be tough to get into the, you know, break in with that, especially if what you do or what could do can now be realistically done by maybe a codex or clog code. Or at least in the next six months, it will. So it's not even like if you're a student today, not only are you thinking about like, geez, I mean, if I'm in marketing, if I'm in digital design, if I'm in these things where Canvas is playing huge roles right now. Lots of things. I was literally just talking to my daughter about architecture yesterday. And I'm not an architect. I have friends who are or whatever. And we were talking about her future and things that she enjoys. And I was like, yeah, I would also have to be careful with architecture because it stands to be reasonable that AI can hoover up all the local codes and figure out very quickly a lot of what might go into an architectural design plan. Now, I'm no shade on any architects out there who are like, you're an idiot, Brian. My point is that's an area where I would go for my daughter. I don't have to worry about her and that decision today. She's not coming out of college for seven years. Let's say six with good behavior, right? So I'm talking about will it affect her job in six years? You've got kids in college right now who came into college with one idea and have probably seen their prospects of a coming out of college with an $85,000, $100,000, $125,000 solid job that was there. And it's poof gone. And that's got to be I can see why you would boo. Now, second point. I love college. I love commencement speech season. I will tell you one of the worst ones I ever saw was one of the ones I attended. It was the year before I graduated in 2000 from UGA in 99. So the year that Amanda graduated, I was at that graduation speech as well because I didn't know. I knew Amanda. But anyway, I was there for a buddy of mine. And it was Ted Turner who recently passed. And Ted kind of just went on a Ted talk. Ted Turner talk. Ted talk. But Ted went on his own. If you remember Ted from the 90s, you remember some good speeches. There's kind of like a Ross Perot thing kind of going on. He never really knew what Ted Turner was going to say. And he went off script. And it was more comical. I mean, I remember it now out of all the things. Also, one of the best ones I ever heard was Agnes Scott College, Kurt Vonnegut. By far, you should go see that one. So that's in my upper echelon. Kurt Vonnegut, Emerson College. My mom was actually at that one for a friend and saw that live, which is super cool. Recently, Eric Church, country singer. Please, please go watch that. This is just out to anybody. Go on YouTube and find Eric Church's commencement speech. I will not give it away. He uses his guitar to talk about the six strings of the guitar and what they could mean about your life. And it may sound silly, but I'm telling you, I got goosebumps. I got goosebumps right now. I immediately sent it to my brother and dad. And in a text, I said, this might be one of the best commencement speeches I have seen in recent time. It's really good. I haven't yet because we were traveling and blah, blah, blah and all things, but I will absolutely have my own daughter sit down and watch it because I just think it's good. It's good. So check that out, Eric Church. Since you gave us all yours, I have to just mention one. Do it. I went to Dartmouth undergraduate and Roger Federer gave the Dartmouth College commencement speech in 2024, just a couple of years ago. Okay. And it's really, really impressive. You'll find it on YouTube. And the main, the core message of that commencement speech was, hey, do you know how, what percentage of matches I won in my career? It's something like 85% of matches that he won. He said, you know how many points? What percentage of points I won? 51%. So he says, you can't focus on the points. Every point is not that meaningful, but you have to be better than 50% in order to get to 84% of matches. And, and that's just a profound insight into the level of focus and persistence that's required to become a champion at anything and to acquire mastery at anything. And that's just one piece of what he conveyed at that, at that commencement speech. So check that one out. That's a good one. Okay. I got to throw one more out in here. Cause Gwen says, what about the army guy who talked about making your bed? That was a really good one, especially in today's time. I don't know who that is. Gwen, I'm sure we could do a quick fact check and find it out, but I do remember make your bed. I do remember that as part of it on there. So yeah. Any, any bet that come to mind? I don't know. Just random ask about commencement speeches, right? I know just thought I'd ask before we moved on for sure. Unprepared for the commencement speech conversation. I get it. It was just Andy mentioned it and I thought, oh, I got, I have to bring this up. I know it's not AI related. I mean, whether Eric church used any AI to help writing the speech, I mean, God, who cares? Who cares? It was really, really well done. Really nails him. He was, he's a guitar player, right? You know, for those that know him or whatever, even if you don't know him, I highly recommend it. It's for, um, it's a UNC, you know, New York University in North Carolina is where he did the speech. And I've, uh, I pulled it up so I can watch it later. Uh, Roger Federer is commenced in speech. Um, there's a New York times post about it actually about a year ago. Yeah. By June 10th of last year. So I want to look forward to watching that one as well, Andy. Okay. Back into AI here. Um, Beth, you want to go next or you want me to share something about Hey Jen? So I have a, I have sort of a weave off of Google IO. Um, there is a research scientist named, uh, Lun Wang, Dr. Wang Lun, Lun Wang. Uh, and he resigned on the 17th and he published the resignation, uh, the, what he, his final blog post in, uh, yesterday and it's about evals. So basically he's saying, um, uh, he's leaving deep mind, incredibly grateful, right? We've seen this as people leave. Um, and then wrote the like, Hey, on my way out the door, this is the parting message that I would like to leave. That's sort of the pattern that we've seen. And basically what he's saying is the thing that I've been thinking about a lot is evals. And basically we're good at evaluating the models that we have. We're worse at evaluating the models we're about to build, right? The, um, and part of what he's saying in this and why I'm interested in this as a conversation is because as we move toward new behaviors, we don't have mod, we don't have evaluation methods to a say that the new behavior is reliably, uh, emerging versus just like a blip, but also how we're evaluating it because evaluations by their very nature are things we do after we see what it is and what the, you know, how you can predict it and that kind of stuff. And, um, and it's really interesting because as I went further into this, I suspect we're going to be borrowing from soft science, right? So the kinds of, uh, circumstances like in ecology, what are the things that happen in a system before the system collapses, right? So what's the temperature or, uh, nutrient in a lake before the algae bloom hits, right? Was fine. And then boom. And those kinds of confounding, uh, circumstances are the kinds of things that I think we're going to go for, go towards. But in the conversation of like, Ooh, these new models are going to blow us away. It seems super relevant that, uh, and we don't know how to evaluate them. Yeah. And the other conversation that I was sort of having, no, I was really having, uh, before the show was a system that I created in college where, uh, because one of the ideas of evaluation is that evaluation almost always be, especially in terms of intelligence, almost always becomes not a, not just an evaluation of the thing you think you're evaluating, but evaluating the model in evaluation, right? So how is it gaming you? How is it understanding what, uh, what the evaluator needs for that? And I did that in college, right? Like I had a whole system. It turned out I learned the material because that was the best way to get the outcome that I wanted, but the outcome was the GPA, the A, the recommendation from the professor. It wasn't learn the material because I'm in college ready to learn the material. Right. I had a specific goal. And I think that that is just something that needs to be a part of the conversation because we do point to these, like, uh, you know, the benchmark tests and say, Ooh, look, it got 89%. And so it's clearly a better model because the other one got 84% and a year ago they got 28%. What is 89% measuring what we were measuring? Like, is it a reliable direction? Yeah. And, you know, does, should we change the model for some of our internal tools? Right. Like, and we, we've seen what a move from 4. Point. Where are we at? 5.5. 4 to 5.5. Let's just use that as an example. Uh, I was like, where are we in Claude? 4.7. Um, uh, but regardless moving even a decimal point on that, you know, that written latest release, I think Amy, you were saying we might see, or maybe Beth, you were saying, whoever was talking about IO yesterday, we might see a, um, a new model from Google, maybe as early as today, maybe tomorrow we'll see what does that incremental increase actually mean to your point, Beth, we're probably going to see the benchmarks, but it really does come back down to how are we, whatever your company is, whatever, how are we using, how are we using this model? And did, did they meaningfully change anything that's valuable to us on it? To me, that would be like, I have a iPhone. I don't know, 11. Right. And it's fine. And I'm deciding this year in September, if I want to move up to the iPhone, what would be 17? Yeah. 17, right. This year. Um, and, and so for me, I may go like, well, it's, it clearly has a better camera. It's if nothing else, it does nevermind all the AI stuff. It's got a better camera on. I'm going to be able to take better pictures. Right. And so I can look at it and go, well, for that reason alone, but then what you get from iPhone 15, iPhone 16, iPhone 17, just like Android models or whatever, the nuances are, are much finer. And so, you know, having an evaluation that says, well, does, does the effort to swap out this model and all of the repercussions that may come with it, is that even worth it? If what the benchmark jump is doesn't have anything really to do with what we do as a company. And so I think you're absolutely right, Beth. Like it's not only evaluations is it's personalized evaluations. Like what matters to you because it may not, what matters to you and what that model does may not matter to other people. And so you have to be able to come up with your own system, just like you did in college of what actually would move the needle for you. But you're right. I think a lot of companies have no idea what that is. And you've said a couple of times in this show, like stock prices don't lie. Numbers don't lie. And those are true. But the stories we tell about them may very well be absolutely not related to what that number means. Right. It may not be predictive. And I feel like that conversation is going to get swept under the rug because people already are saying, what, why? Why are we building capabilities as quickly? Like that is not a requirement. Wouldn't it make sense if we were slowing things down, if we understood what was happening? If we had more of a sense of like, if it goes bat crazy, are we going to be able to do something? Right. Is it possible to unplug? And we say on this show regularly, it's actually not possible to unplug all of AI and go back. Yeah. I don't think you'd want to. I mean, you know, all of it. I don't think you'd want to. But I suppose there's reasons why people might want to go back a model or two. That I see where people, you know, they have preferences or something was just more stable for the stuff they had or their prompting was all geared towards that model. And the new model, you know, with Claude, I think, you know, it's it's much more literal. At least that's what I've read from people. I don't know that I have a stance on that, but I hear a lot of people like TikTok and other people I listen to say, oh, Claude, 4.7 is more literal than previous versions. And that would definitely impact how you write your prompting and use agents. And nowadays it's so much more than, oh, I just I just fix a couple prompts. You know, with agents, everything has gotten more complicated. Well, and that's sort of the convergence that we're starting to see. Right. Is that because Anthropic built so that it could capture some of OpenAI's codex personality and codex built so it could capture some of Anthropics? Right. So now we're closer together than we were before. Or is there also a pattern because we're seeing some of that, that there is enough similarity that the models are starting to converge in or be less detectively different in the way, in the capability, in the way that they share, in the way that they create information, in the way that they create text images, those sorts of things. I think. Yeah. It's all good points. You know, still, still moving, still looking at what's coming out. There is also related to that. There's a there are a couple of videos that have been put out about Google Omni, which Andy talked about yesterday and its video capabilities as explaining science concepts. And this is one of those things that is very, very cool. Right. Like this is photosynthesis. We we are a video maker. Weird. So look, we have just generated this video about photosynthesis. And that's something we can go. We can check. I hope people go and check. Right. They're putting it out. Clearly, some people have gone and check. But it's one of those pieces where I think we're getting into blink territory. That book that was written by Malcolm Gladwell. Thank you, Brain. Um, uh, because it's being presented in such a polished way. If we check photosynthesis and we check the next four concepts, do we stop checking at the fifth? Hmm. Right. Because because people are checking the physical like this came from this paper when you see those little pills in the return of the, um, uh, in, in return of like chat GPT's answer or perplexity's answer. Are we now also just going to have this memory of how this process looks? Because I saw a video about it and there was a very authoritative voice that I heard. And now I have a total misunderstanding of what that concept is. Well, that's a good weave for me to bring up, uh, one of the topics I wanted to bring up today, which was, uh, Hey Jen, which I have not talked about forever on this show. Um, however, I will say that Gareth mentioned in the comments that he uses agent Hey Jen every day. So Gareth, next time you're on the show, um, obviously feel free to add to, to this and how you're using it. Cause I'm not using it day to day, but I was curious because I got an email today. What's the reason I'm even talking about? Well, I got an email from Hey Jen saying today we ship custom motion for avatar five. So you can direct every gesture expression and glance. Uh, and then they gave some examples like look at the camera, lean in, stay serious, uh, smile throughout and use open hand gestures. Another one was count on fingers when listing the three benefits. Another one was thumbs up at the end. It says the same script can be polished because, uh, can become a polished business update, a high energy social video or calm, sincere moment, all directed by the performance. So I was curious. So I was like, well, I, you know, I haven't, again, haven't used Hey Jen in quite a while. In fact, about to share my screen and, and show you just by what I look like on the thing. This is a firm few years ago, December 1st of 2023. Um, so not quite three years, two and a half years ago. So, and, uh, you can see a, uh, a much, uh, shorter haired, uh, thinner Brian in this, uh, and not as tan. So I guess that's a good thing. I got one thing going for me and then I'll show you what I did today, literally this morning before the show, uh, just as an example of where, where we are. And there's definitely an improvement here, but after I show this, I think there's a bigger conversation. Maybe when Gareth comes in as well, uh, we could talk about it, but let me just share what I made a video of, um, again, two and a half years ago, uh, in December. And then let me show you what I did this morning. Uh, let's see, make sure the audio is on. Yep. Okay. Perfect. Okay. So two and a half years ago, this is what the video now you, this is something, if you're listening, I don't think the listening is really going to say it. So, you know, look back at this show at about the 40 minute mark, 38 minute mark, and you can go watch this part of it. If you happen to be listening to us and not on one of our video platforms. Um, but let me just show this part first. Hey, Jen is getting better and better at its video app. This video was created by me type words and trained off of a two minute sample video. I recorded the idea is that I can create new videos on various topics that do not require me to actually set up my camera and report. Now, will people be convinced this is actually me? Probably not yet. So I wouldn't use this for personal notes, but for a quick explainer video, sure. This will do for now. Okay. So it's not great by today's standards, which we would expect. And I had to go back and look, it looked like I was using, uh, what they call avatar three or two. Geez. Now I can't remember. I think it was three. So they're now on five. So now I want to, uh, jump over to two different options that we have. Before you jump, let's just say that what people were seeing was Brian's face. And that voice that you just heard was supposed to be Brian's voice. Yeah. Wow. And so, and that's a good example on this one, because let me see, let me over the past. This one, it was supposed to take my voice and do like from the 15 second clip. This does not sound like me at all. Um, but I would say, uh, look less on or listen less to the way I sound, which is me. Then one I'll show you after this, I recorded the audio for. So it's exactly my voice. And it, and then what I have is essentially just a lip sync. What I would pay more attention to though, is how much smoother without leading the audience, how much smoother this looks. And then we can have a discussion about whether this still looks real or not. Cause I think that's the bigger conversation, but let me just make it a little bit bigger on the screen. Let's make sure the volume's up. Okay, here we go. Over the past two and a half years, myself and the daily AI show co-hosts have done more than 700 live episodes, each one roughly an hour long talking about AI. I have also written 99 newsletters about AI. And what follows is what I wrote in number 100 yesterday. I thought I would share it with my audience here on LinkedIn over the. Okay. So before I move on to the third one, which is just again, me recording the audio myself, the intro, the LinkedIn intro posts I had when I shared the number 100, what 700 live shows taught us about AI. I've talked about that on the show, by the way, I just posted that in the daily AI show community. So if you want to go read that for free, you can by going to the daily AI show community.com and it's up here on the screen if you need it. And that'll take you over to our free Slack community where you can enjoy it. Now, one thing I would say and get your guys feedback here is there's definitely more emotion, right? There's more my head bobbing forward and my head kind of moving around and stuff. It definitely looks more. It's definitely an improvement over what two and a half years ago was. And now we can look just finally at the. Wait, but I want to add a comment about the quality of the Hay Jin range of, of our articulations of your face, your head, your body and so on. They seem to be on a loop to me. As opposed to you in live performance, you have a little more subtle and nuanced variations in what you do that are a function of maybe just even the pauses that are happening in the generation of your speech. And so you have these associated gestures and so on. And so there's a wider range of physical performance than is represented in the Hay Jin system. So in the Hay Jin version of you, it's like, oh man, he's just going back and again to that same head bob. I can see that same head bob happening. So I wanted to say that while it's better, it's still kind of clunky and robotic to me if I really look at it. I agree with you. Let's go ahead, Beth, before I will play this last one. I just want to, because this will be interesting to see in the last one too. One of the things that I think hasn't been taken into account is the progressive flush or paling when we get excited about something that we're talking about, right? So Brian, your cheeks stay the same red in the whole thing, right? And I don't think that AI has started to take that. They're like, ooh, well, when you're sad, your eyes do this shape, right? And your voice does this pattern. But we have all kinds of tells. I agree with you, which is why this hasn't passed the uncanny valley for me. I think either. I don't want to include y'all, but that's what I think. Now, this is my real voice. So nothing at all. I just want to be clear here. The last one, Hay Jin tried to quickly mimic my voice and we heard that it didn't do a good job. This is me just literally recording it. But what Hay Jin did was produce the video to go with the audio. So now we can watch this one really quick and watch, again, the head movements and things like that. Over the past two and a half years, we've done more than 700 live daily AI show episodes. We have covered launch days and benchmark contests, boardroom fears and investor hype, safety disputes and GPU shortages, roadblock demos and open source surprises. We have watched companies present AI as a product, platform, threat, utility, co-worker, creative partner and infrastructure layer, sometimes shifting positions within the same quarter. After all of those shows, one lesson stands out. Now, this one is decidedly better, but I think it's better back to your point, Andy, because Hay Jin is picking up on my natural pauses and my highs and lows in my voice. And instead of going more in that sort of loop, because there's not enough variance in the voice it created, there seems to be a better video of this. I will tell you guys, I'm pretty impressed with this. I think, you know, there's maybe not enough blinking eyes and the way my head is positioned up could could have been because of the way I was trying to read the script or whatever. There's a lot of like weird introduces. This all comes back, though. And this is the conversation. And again, I really want to hear what Gareth is building with it, because I'm not saying there's not a good use case. And, you know, back when Aaron was on the show, he used to talk about Hay Jin all the time and show us the latest and greatest. But what I remember is Hay Jin wasn't wasn't necessarily cheap. And yes, this is just me playing around with whatever they gave me access to. I didn't want to buy anything. So fair. Maybe I don't have the latest and greatest from Hay Jin. But it still goes back and says, well, if I have to record the audio, which, by the way, I had to do three times because I kept flubbing on the last word in the last, you know, that kind of thing. OK, no big deal. This is only like 25 seconds long. It didn't really take a whole lot of time. And yeah, I suppose I didn't have to be on camera ready, although obviously I look exactly like the video. So that wouldn't have been an issue for me. And so I really get down to like, OK, but so where did where did this help me? Because if it's not really a replacement for me, the only way that I see this being valuable is if you just need like a talking head video and it has to just be good enough. But nobody is really thrown that it's AI. And there are use cases for that where I'm fully aware that the video I'm watching is an AI avatar, but it doesn't matter for the situation. And that's fine. It's only really when it's trying to be passed off as human and it's not. And that gets into some, you know, I want to be I want to be told what I'm looking at. I don't want to be like it was like somebody's trying to, you know, fake me out on it. So I don't know. What are your guys takes? I mean, this is this is the latest and greatest. Now, I will tell you, I didn't have the option like the email was saying about doing certain like click or snap my fingers or whatever it was. It's weird. It was like there and I could do it, but it didn't bring it into the video. That could just be because I was using like their free credit, whatever, whatever. So fair there. But I'm still trying to figure out where I'm still trying to figure out where agent is a really good use case because as good as it's gotten better, as much better as it's gotten in two and a half years, which I am. No, you can see the difference how much better the technology is. I still have a hard time finding the use case. I've got one for you, but I want to see what Andy's going to say. I wanted to say that the the quality may be partly a function of how much training video you give it. So I'm thinking that this demo you did is is based on just a very short clip of video that you provided today with that outfit on. And and by time, image in this case, by the way, it was just just. Just a new. OK, so I think that that agent actually has the potential to get much, much better representation of you and the range of your gestures if you give it a lot more video for training. But that's expensive. Like you have to. I think in the agent subscription world, you have to you have to pay a lot of money in order to feed a lot of information about you. But if anybody I know that Sabrina Romanov has has done hey, Jen a lot and uses it in the in the production of of some of her videos and some of the videos that are hey, Jen of her like hey generations of her are, in my view, indistinguishable from her life. But in part, she is very clear. This is a hey, Jen video. She's not trying to pass it off. Ryan, I can see I think that there's a there's a time limit and it may be like six seconds, seven seconds of the full body like you're holding the whole screen. And I can imagine you would do it for the trainings, right? You literally go places or do it online trainings for people for your company. I can imagine you do like like, hey, this is my personalized greeting to your company. And what we're going to talk about is this. And then it shrinks down into a tiny bubble where I'm not paying attention to the same kind of things like and then that could be done at scale. From scaled at scale. Yeah. I see what you did there. Smooth. The I look, you're right. The tiny bubble in the lower left corner of a video, whatever. Who cares? You know, is it better than not having avatar or be a comic avatar like, you know, sure. You know, to do personalized training for people. I get that for me, at least. And this is just me. I still find it to be a time suck and a cost over what it would take for me to just simply do the video. The video. But maybe I'm just not at the volume that would tip the scale there where I go, well, I can't. To your point, Beth, I have a situation where maybe I have to. I don't. I'm saying hypothetically where I need to create 50 personalized versions of this. Sales outreach would be a good version of that. You know, is me doing that. And let's say I'm a sales rep and I'm doing outreach. And I do want to basically spin these up and get them out the door as a quick, you know, a little video or whatever. But then to me, you just run right into the same old problems that you had before, which is now you've got to, you know, an email with a video on it and it's clicking a link and a lot of that's going to get blocked by a lot of process. I don't know. I really don't know. I just think, except for like maybe some of the limited use cases like you guys are saying and you brought up back, I just don't know exactly where this is for. I'm not knocking that the technology is impressive as hell. I just don't know where I would, I guess I would use it. I'm sure I have a limited view on this. I'm sure I do. Because what we're also talking about is the cost loss factor. Is it worth your time save for someone to look at it and say, I think that's a I. I watch him on the Daily Egg Eye show and he's much more expressive in his face. Right. Those kinds of things. Like how long did it take you? It took you three minutes to record the intro twice. Right. Those kinds of things. And the other thing, let's just warn everybody, like when you are considering doing this in your company, Carl used to talk about this way back. Make sure that you don't give your company the perpetual license to use your likeness on the training, because that's what this technology is for, too. Yeah. Oh, by the way, I will say this because one of my old videos was me speaking French, not literally. It was a joke, too. We just started the show, but people were laughing and saying I sounded French Canadian, not like Parisian French. And people were laughing at it. Those who could speak French, they're like, it's not bad. But like, it's not like it's not right. You know, either. It's French Canadian. And it was supposed to be like France French. Right. So, you know, dialects and stuff. But anyway, you know, translation, I do think I want to I want to put that out there on the thing. That's an area where probably people don't care as much that it's an AI avatar. They know I'm not speaking French. And so it's a lot more obvious, but it still has that personal feel to it. And so I do still see value there if the heads of the company or whatever are emerging into a new market and they want to create something in that in that language. As long as everybody's on the up and up, that that's what it is. But it allows you to get a little bit of my, you know, expressive behavior and my voice. But you're hearing it in French or German. I see value there because I think everybody's in on it and knows what they're looking at. They're like, oh, they went ahead and took the extra time to create a CEO message in German. Great. I know what I got. You know, I'm not I'm not trying to be full. Nobody's trying to pull the wool over my eyes. I don't think the CEO now speaks seven languages. Just they're using a tool to do that. I'm aware of it. We're all good. And I think that goes back to all these conversations about like, what is your intent? Are you trying to fool people into believing that you recorded this live? And then in fact, it is you. We talked about it with the cop. There's a lot of booze and pushback to that. But people don't want to be do not want to be fooled or perceived as like they don't know what's going on. So I would just say to anybody that's that's working on this stuff, like think about your use case. But if your use case is, oh, this will allow me to not have to get on camera. And I can put out lots of now personalized videos and in the intent is to pass it off as I made the video. I would say careful, careful, careful, careful there, because you might get more blowback on that and reputation damage. And the time saved is certainly not going to save any money. I mean, time is money, but agent is going to cost you money to get it right to your point with like Sabrina. So anyway, I thought it was interesting. I saw the email. I thought, well, let's see. Let's see where we're at, because we just I haven't personally talked about it on the show. But yeah, I mean, impressive, right? What they what they've been able to do over the last two and a half years. All right. I've got three quick news items I want to throw out before we close. First of all, OpenAI, as we spoke about on the show, had taken the CEO position of Sam Altman and paired it with another CEO level position for Fiji Simo over product. Then there were some announcements about her needing to take away some time away from OpenAI for medical reasons. And then we had heard a lot from Fiji Simo recently. And today there's news that OpenAI is consolidating product under Greg Brockman. So I think that my conclusion, and I'm speculating, is that Fiji Simo is out for a while. And so Greg Brockman is taking over the product management. And one of the things that they're doing in that product region of the application of all the tools that they build is they're merging chat GPTN codecs. And so this seems to be an implication that they view coding and developer use as converging together with the way consumers are using it. So that, you know, coding will be generated for responses to consumer queries and vice versa. Developers will, you know, want to have more consumer-like interfaces. And so they see an alignment between those two. Okay. The second thing of three is that NVIDIA has kind of given, you know, a kudo and an endorsement of the Hermes agent. And what they did is they published a guide for running Hermes agent locally on its hardware. So they threw its weight behind new research, the writers of Hermes agent. And they provide a detailed guide for how to run a local agent on NVIDIA, DGX Spark, or just any RTX PCs. And they're pairing it also with Alibaba Quinn 3.6 for the open weight model underneath that. So they're inviting you to take Hermes agent instead of Nemo claw, right? Which was they immediately after Open Claw was released, they, you know, put out Nemo claw, which is their version of Open Claw. But now they're also saying, hey, use Hermes agent because Hermes agent is really, it's getting the GitHub stars. It's, you know, getting a very large number of downloads and so on. They got 140,000 GitHub stars already in three months. So anyway, Hermes agent, I know Beth, you're using it. And I'm planning to implement it eventually. So NVIDIA is endorsing that in that way. Finally, there's news that's not so recent news, but it's just starting to come to my attention. And there's a couple of different newsletters today that referred to this. And that is that Mythos, the mythical high-powered model from Anthropic that's being used in the cybersecurity space, is actually better than codecs in the coding benchmarks. So specifically, Mythos is getting 94% on the SWE Bench verified benchmark, while GPT 5.5 as used in codecs gets only 83%. And then on this much harder coding application benchmark called SWE Bench Pro, Mythos preview is at 78%, and GPT 5.5 is at 59%. So I just thought that was an interesting comparison that, you know, put that in your pipe and smoke it if you've been declaring codecs to be the winner. It's true today. But when Mythos or a CLAWD 4.8 comes out that has Mythos capabilities baked into it for coding, there's going to be another jump. But, of course, behind the scenes, OpenAI is certainly coming up with 5.6, which will probably be at parity with that. So it's just a tit-for-tat kind of thing. But right at the moment, CLAWD and CLAWD code, you can rest assured that what's behind the scenes and will be implemented in CLAWD code eventually is going to be, you know, right at the frontier, if not beyond the frontier models that are playing in coding right now. Absolutely. And I don't think we've said this yet, but speaking of Altman and tit-for-tat, Elon lost the court case. You're right. And you might say that we all lost the court case because we know, or won, we know so much more about what was happening. Lots of, I don't know, infighting came to light, those kinds of things. And, like, and basically the decision was it's too late to decide, right? The decision was statute of limitations has passed and you did not make the case that we should ignore that the statute of limitations has passed. So, yeah. Okay. Well, we'll be back tomorrow. Jumi will be here. And I think Carl is going to be here tomorrow. And I hope everybody comes back and has a great time then. Go build with AI. And I think we're about to wrap. Andy, final words? Nope. All right. I'll skip a wrap. So, again, IO is today. Watch the, watch the keynote. I don't know that we're going live, but if you hit the bell, you'll get notified if we do. All right. Bye. Bye.